Refactor Varimp

DESCRIPTION

@@ -45,7 +45,6 @@ Imports:
     methods,
     pbapply,
     ggplot2,
-    digest,
     lattice,
     gridExtra,
     data.table,

NAMESPACE

@@ -38,12 +38,14 @@ importFrom(caret,varImp)
 importFrom(data.table,.SD)
 importFrom(data.table,as.data.table)
 importFrom(data.table,data.table)
+importFrom(data.table,dcast.data.table)
+importFrom(data.table,merge.data.table)
 importFrom(data.table,rbindlist)
 importFrom(data.table,set)
+importFrom(data.table,setcolorder)
 importFrom(data.table,setkeyv)
 importFrom(data.table,setnames)
 importFrom(data.table,setorderv)
-importFrom(digest,digest)
 importFrom(ggplot2,aes)
 importFrom(ggplot2,autoplot)
 importFrom(ggplot2,geom_bar)

R/caretEnsemble.R

@@ -77,7 +77,7 @@ getMetric <- function(x, metric = NULL, return_sd = FALSE) {
 #' @param metric a character string representing the metric to extract.
 #' If NULL, each model will return the metric it was trained on.
 #' If not NULL, the specified metric must be present for EVERY trained model.
-#' @impotFrom data.table data.table setorderv
+#' @importFrom data.table data.table setorderv
 extractModelMetrics <- function(ensemble, metric = NULL) {
   stopifnot(is.caretEnsemble(ensemble))
   model_metrics <- data.table(
@@ -117,28 +117,25 @@ summary.caretEnsemble <- function(object, ...) {
   print(extractModelMetrics(object), row.names = FALSE)
 }
 
+#' @title Calculate the variable importance of variables in a caret model.
+#' @description This function wraps the \code{\link[caret]{varImp}} function
+#' from the caret package. It returns a \code{\link[data.table]{data.table}} with importances normalized to sum to 1.
+#' @param x a \code{\link[caret]{train}} object
+#' @param model_name a character string representing the name of the model
+#' @param ... additional arguments passed to \code{\link[caret]{varImp}}
+#' @return a \code{\link[data.table]{data.table}} with 2 columns: the variables and their importances.
+#' @importFrom caret varImp
+#' @importFrom data.table data.table
 #' @keywords internal
-# This function only gets called once, in varImp.caretEnsemble
-varImpFrame <- function(x) {
-  dat <- do.call(rbind.data.frame, x)
-  dat <- dat[!duplicated(lapply(dat, summary))]
-
-  # Parse frame
-  dat$id <- row.names(dat)
-  dat$model <- sub("\\.[^\n]*", "", dat$id)
-  dat$var <- sub("^[^.]*", "", dat$id)
-  dat$var <- substr(dat$var, 2L, nchar(dat$var))
-
-  # Parse intercept variables
-  dat$var[grep("Inter", dat$var, fixed = TRUE)] <- "Intercept"
-  dat$id <- NULL
-  row.names(dat) <- NULL
-  dat <- reshape(dat,
-    direction = "wide", v.names = "Overall",
-    idvar = "var", timevar = "model"
+varImpDataTable <- function(x, model_name, ...) {
+  imp <- caret::varImp(x, ...)
+  imp <- imp[["importance"]]
+  imp <- data.table::data.table(
+    model_name = model_name,
+    var = trimws(gsub("[`()]", "", row.names(imp)), which = "both"),
+    imp = imp[["Overall"]] / sum(imp[["Overall"]])
   )
-  row.names(dat) <- dat[, 1L]
-  dat[, -1L]
+  imp
 }
 
 #' @title Calculate the variable importance of variables in a caretEnsemble.
@@ -148,47 +145,39 @@ varImpFrame <- function(x) {
 #' importance for each model is calculated and then averaged by the weight of the overall model
 #' in the ensembled object.
 #' @param object a \code{caretEnsemble} to make predictions from.
-#' @param ... other arguments to be passed to varImp
-#' @return A \code{\link{data.frame}} with one row per variable and one column
+#' @param ... additional arguments passed to \code{\link[caret]{varImp}}
+#' @importFrom data.table rbindlist dcast.data.table merge.data.table setorderv setcolorder
+#' @return A \code{\link[data.table]{data.table}} with one row per variable and one column
 #' per model in object
-#' @importFrom digest digest
-#' @importFrom caret varImp
 #' @export
 varImp.caretEnsemble <- function(object, ...) {
-  # Extract and formal individual model importances
-  # Todo, clean up this code!  Make varImp.caretList
-  coef_importance <- lapply(object$models, caret::varImp)
-  coef_importance <- lapply(coef_importance, function(x) {
-    names(x$importance)[1L] <- "Overall"
-    x$importance <- x$importance[, "Overall", drop = FALSE]
-    x$importance
-  })
+  model_names <- make.names(names(object$models), unique = TRUE, allow_ = TRUE)
 
-  # Convert to data.frame
-  dat <- varImpFrame(coef_importance)
-  dat[is.na(dat)] <- 0L
-  names(dat) <- make.names(names(coef_importance))
+  # Individual model importances
+  # TODO: varImp.caretList should be a separate function
+  model_imp <- mapply(varImpDataTable, object$models, model_names, MoreArgs = list(...), SIMPLIFY = FALSE)
+  model_imp <- data.table::rbindlist(model_imp, fill = TRUE, use.names = TRUE)
+  model_imp <- data.table::dcast.data.table(model_imp, var ~ model_name, value.var = "imp", fill = 0.0)
 
-  # Scale the importances
-  norm_to_100 <- function(d) d / sum(d) * 100.0
-  dat <- apply(dat, 2L, norm_to_100)
+  # Overall importance
+  ens_imp <- varImpDataTable(object$ens_model, "ensemble")
+  ens_imp <- data.table::dcast.data.table(ens_imp, model_name ~ var, value.var = "imp", fill = 0.0)
 
-  # Calculate overall importance
-  model_weights <- coef(object$ens_model$finalModel)
-  # In the case of 2 classes each method will
-  # have only one coef associated.
-  # The names of the weights keep the order of the
-  # models in the ensemble
-  names(model_weights) <- names(object$models)
-  model_weights <- model_weights[names(model_weights) %in% names(coef_importance)]
-  model_weights <- abs(model_weights)
-  overall <- norm_to_100(apply(dat, 1L, weighted.mean, w = model_weights))
-  dat <- data.frame(overall = overall, dat)
+  # Use overall importance to weight individual model importances
+  model_imp_mat <- as.matrix(model_imp[, model_names, with = FALSE])
+  ens_imp_mat <- as.matrix(ens_imp[, model_names, with = FALSE])
+  overall_imp <- data.table::data.table(
+    var = model_imp[["var"]],
+    overall = (model_imp_mat %*% t(ens_imp_mat))[, 1L]
+  )
 
-  # Order by overall importance
-  dat <- dat[order(dat[["overall"]]), ]
+  # Merge overall importance with individual model importances
+  imp <- data.table::merge.data.table(overall_imp, model_imp, by = "var", all = TRUE)
 
-  dat
+  # Order and return
+  data.table::setorderv(imp, "overall", order = -1L)
+  data.table::setcolorder(imp, c("var", "overall", model_names))
+  imp
 }
 
 #' @title Plot Diagnostics for an caretEnsemble Object